Updates to Congressional Speech Acts on Twitter - Shapiro, Matthew A.; Hemphill, Libby; Otterbacher, Jahna ...
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
University of Michigan Deep Blue deepblue.lib.umich.edu 2018-08 Updates to Congressional Speech Acts on Twitter Shapiro, Matthew A.; Hemphill, Libby; Otterbacher, Jahna http://hdl.handle.net/2027.42/145592
Updates to Congressional Speech Acts on Twitter
Matthew A. Shapiro, Illinois Institute of Technology*
Libby Hemphill, University of Michigan
Jahna Otterbacher, Open University Cyprus
Abstract
Over the last several years, research on our elected officials’ use of social media as a political
communication platform has greatly increased. While the bulk of social media-related research
focuses on elections, social media-traditional media connections, or the effect of politicians’
social media communications on people’s attitudes and opinions, the present study shows how
members of the U.S. Congress use Twitter to engage in a range of speech-based actions.
Examples of these speech-based actions include narration about one’s day or recent events,
providing information in the form of online or offline information, and positioning for/against
policies and other politicians. In terms of outcomes, this paper provides updates regarding
gender, chamber, and party-based differences. Second, based on the assumption that speech
acts now occur in hybrid form, this paper examines how polarizing political communications are
couched in more subdued formats. Third, a set of recommendations is provided to help
journalists and citizens identify these hybrid speech-based actions before making a potentially
misinformed retweet or comment. In this way, the function of Twitter use by elected officials is
further explained and our understanding of Twitter’s role in U.S. political communication is
further deepened.
~ Prepared for the 2018 American Political Science Association Annual Conference ~
* Corresponding author: shapiro@iit.eduIntroduction
The present study approaches politicians’ social media use by revisiting earlier findings that
show how members of the U.S. Congress use Twitter. In this way, we provide both an updated
and deeper understanding of Twitter’s role in U.S. political communication. Based on a dataset
of 1.2 million tweets posted by 529 Twitter accounts, we discover here that much has changed
over the past several years: gender, chamber, and party-based differences have all reversed in
terms of predicting Twitter post frequency. Based on the results of an automated classification
process for what we call “speech actions,” we relax our earlier assumption that communications
on Twitter can be broken into specific, single speech actions rather than multiple speech actions
simultaneously. Our analysis of these “hybrid” speech acts shows that members of Congress
are now simultaneously providing information and positioning via Twitter, and there are
significant party-based differences. Gender-based differences are nonexistent with regard to
predicting hybrid speech acts, and congressional chamber-based differences are significant in
only one instance.
The inclusion of hybrid speech acts is crucial for the corpus of research on online political
communications. They are expected to be much more effective at obscuring politicizing content
and couching it in non-politicizing rhetoric. In other words, members of Congress are using
Twitter by embedding positioning statements within other speech acts such as providing
information, requesting action, and thanking. This nuanced method of providing information
invokes theories of the politicization of information and undoubtedly increases the information
asymmetries between members of Congress and the general public. It also has implications for
the traditional media, which relies extensively on Twitter as a primary source of information, if
journalists lack critical literacy skills.
In the following pages, we establish the literature central to this project, particularly the
linguistics of Twitter and how Twitter content can be politicized. We then outline the
classification process, largely building off of Hemphill et al. (2013), and attempt to answer a
number of research questions with our panel dataset of congressional tweets. Following a
discussion section, we conclude with suggestions for journalists as well as general citizens
when encountering varying degrees of positioning in congressional speech acts on Twitter.
Related Work
We recognize that social media is but a part of politicians’ broader information-sharing
strategies, and we build on the literature focusing on how elected officials communicate using
the traditional media (Cook et al., 1983; Edwards & Wood, 1999; Entman, 2007; Kedrowski,
2000; Lee, 2009) as well as online (Gentzkow & Shapiro, 2011). Our focus is not simply to
determine whether and when Twitter rates for members of Congress rise and fall (see, for
example, Chi and Yang (2011) and Straus et al. (2013)). Rather, we are more interested in the
composition and intent of political communications from our elected officials.
1The Linguistics of Twitter
There is not a large literature suggesting that language behaviors in general have evolved over
time on Twitter or other social media. Rather, the bulk of Twitter-oriented research has focused
on changes in linguistic patterns that are then attributed to social relations (i.e., language style
being modified and/or influenced as a function of one’s relations with others). This is consistent
with a foundational literature on linguistic variation holding that 1) people’s linguistic style is
largely predicted by their demographic attributes (e.g., gender, age, where they grew up, etc.),
2) people may change their style as a function of the addressee -- who they are addressing (i.e.,
their audience) and what relationship they have with the addressee (Labov, 1966). Others
employ a social identity approach to explain why people adapt their language patterns, with the
underlying drive being social acceptance from others (H. Giles, Coupland, & Coupland, 1991).
For instance, people of lower social status will adapt to those of a high status, stylistically. Allan
Bell (1984) provided a theory of language change even in the absence of direct interaction
between people (which is an assumption for Labov [1966]), claiming that each medium has a
“house style; e.g., the linguistic styles of newscasters (Bell 1982). This explanation is structured
around audience design in that communicators (in this case people who write for the media)
may not have direct, frequent communication with others, but they do have an idea of who their
audience is based on infrequent contact and/or mass media contact (1984).
Among those studies focusing exclusively on Twitter, Marwick and Boyd’s (2010) research on
the “imagined audience” resonates with Bell’s notion of a house style resulting from audience
design. Their qualitative study of Twitter users revealed that people do largely imagine their
audience on Twitter and design their behaviors as a result of that imagined audience, even
though the audience actually consists of a diverse range of people, known and unknown to
them. Employing a computational approach, Hu et al. (2013) then confirmed that the stylistic
features of tweets differ significantly from the style typically use on other media (e.g., SMS, chat,
email, blog corpora). Twitter’s house style is strikingly more formal than what might be expected
in that sentiment is more positive than negative, less dynamic over time, and exhibits relatively
less usage of slang (Hu et al., 2013).1 Further, members’ language change can predict their
stage in the lifecycle in the online community; when one becomes less accommodating to the
linguistic style of the group, it is a typical indicator of losing interest and abandonment of the
community (Danescu-Niculescu-Mizil, West, Jurafsky, Leskovec, & Potts, 2013).
Early descriptions of Congress suggested politicians use Twitter mostly to provide information
(Golbeck, Grimes, & Rogers, 2010; Hemphill, Otterbacher, et al., 2013). In a parallel vein,
research has examined the specific content or action of tweets, identifying the following
characteristics: attacking, campaigning, mobilization, issues, media, and user interaction
(Evans, Cordova, & Sipole, 2014; Haber, 2011); providing information, requesting action,
positioning, thanking, narrating (Hemphill, Otterbacher, et al., 2013); direct communications,
personal message, activities, information, requesting action, fundraising (Golbeck et al., 2010);
informational, organizational, policy, and attack/negative campaigning (Gainous & Wagner,
1
Some of this stylistic variation has been attributed to power differences between interlocutors
(Danescu-Niculescu-Mizil, Gamon, & Dumais, 2011).
22014; Granberg-Rademacker & Parsneau, 2018); campaign events and locations, information,
personal, attacks, presidential candidates, replies, calls to action, incumbent business,
fundraising (Frechette & Ancu, 2017); position taking, press or web links, district or state
activities, official congressional action, personal, and replies (Glassman, Straus, & Shogan,
2009). This compilation of tweet qualities clarifies a consist focus on campaigning, sharing
information, and, most importantly, politicking.
Within this class of communications, campaigning impacts politicians’ communications in a
unique way, but there was not a marked change in strategy among congressional candidates in
the 2016 election (Druckman, Kifer, & Parkin, 2017). Consistent with the extant literature on
campaign strategy (see, for example, (Druckman, Hennessy, Kifer, & Parkin, 2009; Graham,
Jackson, & Broersma, 2014), newcomers challenged incumbents, and incumbents focused on
policies (Frechette & Ancu, 2017). As well, most of the effort in 2016 was to convey information
rather than mobilize supporters or fundraise (Frechette & Ancu, 2017), although challengers
may have tweeted more frequently to blunt the institutional advantage of incumbents (Evans et
al., 2014). We acknowledge that campaigning plays a significant role in how language is used
on Twitter, but our focus accounts for the entire history of congressional Twitter use rather than
singular events such as elections.
Together, the extant research on the house style on Twitter and the common uses of Twitter by
politicians suggest the following research question:
RQ1: In what ways has Congress changed its speech acts distribution over time? Has
the “house style” changed?
Twitter’s Political Communication Potential
We assume that political communications are rooted in the notion of a frame in a
communication, which refers in the case of Twitter to words (i.e. hashtags), phrases, or images
(i.e. forwarded photos) that highlight certain considerations toward a politician, policy, or issue
(Druckman, 2001). How these frames are both constructed and received are both relevant. On
Twitter, for example, there has been considerable research about how members of Congress
clarify their interest in specific policy issues through the use of hashtags (Cunha et al., 2011;
Hemphill, Culotta, & Heston, 2016; Huang, Thornton, & Efthimiadis, 2010; Shapiro & Hemphill,
2017). For the present study, we invoke Hemphill et al.’s (2013) model of Twitter-based speech
acts. Presented in Table 1, these five speech acts and their definitions were derived from an
iterative hand-coding and automated-coding process.
TABLE 1 HERE
Among the five speech acts identified by Hemphill et al.’s (2013) model, we highlight the fact
that positioning is a politicizing frame and that providing information as a news-update frame.
However, we offer an update to the existing research on Twitter-based political communication
by recognizing the potential for combinations of frames to occur. We focus particularly on
hybridized speech acts involving the conflation of positioning and providing information with
each other as well as with the remaining speech acts. The presence of hybrid speech acts
would indicate a subtle but shrewd framing strategy by members of Congress. For example,
3conflation between positioning and providing information speech acts could provide legitimation
for political content by providing URLs and directing people to outside sources of information.
Similarly, conflation between positioning and thanking could emphasize the compassionate
qualities of politicians over attempts to portray themselves as political leaders. To our
knowledge, these nuanced approaches to political communication on Twitter have not been
addressed in the existing literature.
We also assume that politicians frame their Twitter posts while continuing to recognize the
significance of their online audiences (Karlsen, 2015; Norris & Curtice, 2008; Williams & Gulati,
2010). Some research suggests that politicians are engaging in one-sided or parasocial
interactions rather than reciprocating, human-to-human interactions (D. C. Giles, 2002) and that
such parasocial interactions are nothing more than a facade of interactivity (Stromer-Galley,
2000). However, and in line with McMillan (2002), we believe that citizens appreciate and feel
more proximate to their elected officials via Twitter given evidence that members of Congress
directly respond to their constituents (Barberá et al., 2014; Barberá, Bonneau, Jost, Nagler, &
Tucker, 2013).2 At the same time, the size of one’s audience is a positive function of media
contact as journalists increasingly draft their articles on the basis of social media-based
information (Hamby, 2013; Parmelee, 2013; Verweij, 2011). Politicians acknowledge that the
media will be the ultimate conveyor of information to the general public (Lieber & Golan, 2011)
and thus use Twitter to communicate political statements and policy preferences to the
mainstream press (Shapiro & Hemphill, 2017). If journalists continue to engage in significantly
less fact-checking of politicians when information originates in tweets (Coddington, Molyneux, &
Lawrence, 2014), and if hybrid speech acts are employed as a framing strategy by members of
Congress, the prospects dim for transparent and accurate journalism.
Thus, the potential for hybrid speech acts to arise suggests a second research question:
RQ2: How is the positioning speech act conflated with other speech acts, and what are
the implications of couching positioning with, for example, thanking or providing
information?
Variance among Members of Congress
We acknowledge the frequent emphasis in the existing research with regard to how Twitter can
be used in different ways by different types of politicians. These differences often focus on
gender, congressional chamber, and party. We know that the behavior of women in Congress
can differ from that of men, namely that women are active in their online engagement (Hemphill,
Otterbacher, et al., 2013; Niven & Zilber, 2001). As well, recent research confirms that female
politicians use the Internet for communication -- including Twitter -- more frequently than their
male counterparts (Evans, Cordova, & Sipole, 2014; Evans, Ovalle, & Green, 2016), which
counters existing research on Twitter use (Hemphill, Otterbacher, et al., 2013). In terms of
congressional chamber, in a study of Twitter use from October 2015 to May 2016 among
2
Twitter-based connections from politician-to-citizen are typically limited to constituents that are
co-partisans.
4individuals on the U.S. House and Senate Armed Service Committees, it was found that
representatives tweet more about legislative functions while senators tweet more about
oversight functions (Buckles, 2017). Regarding party-based differences, which are likely to have
the greatest implications if polarizing rhetoric differs between Democrats and Republicans, in an
analysis of Twitter use by U.S. senators, it was found that partisan rhetoric is more likely to be
employed by Republicans (Russell, 2017).
These conflicting findings about differences across gender, chamber, and party suggest a third
research question:
RQ3: How do members of Congress differ in terms of how they use Twitter? Specifically,
what differences can we observe between men and women, between chambers, and
between parties?
Methods
The principal updates to Hemphill et al. (2013) offered in this paper are reflected in both the
nature of the data and in our analytical approach. The entire dataset was recollected, and the
existing model in (Hemphill, Otterbacher, et al., 2013) was used to label speech acts. Our full
dataset includes 765,626 tweets from 414 accounts, representing 77 percent of the members of
Congress in 2016. To collect Twitter handles, we used a crowd-sourced list of official Twitter
accounts for members of Congress from the unitedstates project.3 We collected tweets using
Purpletag’s collect functions (Hemphill et al., 2016). Twitter’s search API returns up to 3,200
tweets for each user, and we therefore have complete histories for accounts with fewer than
3,200 tweets but the most recent 3,200 tweets for accounts with more than 3,200 tweets. The
earliest posting of a Twitter post in our dataset is February 14, 2008, and the latest posting was
made on February 15, 2016.
Automated Classification
The classifiers developed in Hemphill et al. (2013) used to label tweets in our data set employ
MALLET (Machine Learning for LanguagE Toolkit) to train and evaluate our classifiers
(McCallum, 2002) (http://mallet.cs.umass.edu/). Using the maximum entropy classifier, the
words appearing in a tweet are its features, and the classifier is a function that maps the
features onto the output classes. “Maximum entropy models begin with the assumption that
uniform distributions are preferred (i.e., assume a 50/50 chance that a tweet is ‘narrative’ or
not). They use training data to learn constraints to be applied to this distribution. Nigam and
colleagues (Nigam, Lafferty, & Mccallum, 1999) report that in many cases, maximum entropy
outperforms Naïve Bayes, however, it does have a tendency toward overfitting in cases where
data is sparse (i.e., when there are only few positive examples of a tweet of a given class)”
(Hemphill, Otterbacher, et al., 2013: 4). The accuracy statistics for our classifier’s performance
are reported in Table 2.
TABLE 2 HERE
3
https://github.com/unitedstates/congress-legislators
5Statistical Analysis
In line with the research questions presented above, our primary focus is the use of Twitter by
members of Congress. We provide direct updates for previous research on this topic and also
consider specific features of the identified speech acts that had been hitherto ignored, namely
the presence of hybrid speech acts. The dataset itself is unique in its panel structure, and the
time parameter is based at the minute level. When structuring the dataset, approximately nine
percent of the tweets were made by individual members of Congress at the same minute,
creating duplicates in terms of the panel structure. We opted to delete these duplicates to
maintain a dataset free of non-uniquely identified observations. We could have alternatively
transformed the time parameter from the minute to the second level and modified duplicate
tweets’ timestamp by one second. We confirmed, though, that there were no significant
differences between the original and truncated datasets in terms of their descriptive statistics,
correlations, and a preliminary pooled data analysis where the time parameter was included as
a control. The statistical output remained virtually unchanged between the original and
truncated datasets. Random effects linear regression was conducted to identify characteristics
driving overall frequency of tweets. Random effects logistic regression was conducted to identify
characteristics driving each speech act.
Results and Discussion
We first establish which types of politicians are in fact using Twitter for communication. To this
end, we present in Table 3 the results of regressing frequency of Twitter posts on gender,
congressional chamber, and party attributes. We observe that the average member of Congress
posts on Twitter more than 837 times. We also observe that females, senators, and Democrats
are much more likely to tweet; in the case of chamber, the difference between senators and
representatives is on the level of 355 tweets. These findings confirm Evans et al. (2016) in
terms of gender but contrast with (Russell 2017) in terms of party (for senators only). They also
differ completely from the parallel analysis in Hemphill et al. (2013), leading us to claim that
Twitter use by members of Congress has evolved and that the “house style” has in fact changed
over the last several years.
We note that the gap between senators and representatives is a function of temporal variation
between chambers, namely the difference in term lengths between senators and
representatives. This and other potential sources of individual and temporal variance will be
subsequently addressed with qualifying tests for random effects modeling.
TABLE 3 HERE
Based on our sample of 765,626 tweets, Table 4 reports two estimates for the predicted
probability that a speech act applies to any given tweet; the left-hand column reports the raw
MALLET scores (i.e., mean probability that a speech act applies) while the right side conveys
the descriptive statistics for these speech acts after applying a 0.50 threshold to the classifier
(i.e., applying a stricter condition for a label to be assigned). In this way, we are able to create a
binary classifier for each speech act under the condition that MALLET predicts at least a 50
percent probability of a speech act occurring. Under this 0.50 threshold, the results are similar in
6terms of these five speech acts’ overall frequency ranking, but the proportion of tweets
represented by narrating, positioning, requesting action, and thanking is more conservative
under the 0.50 threshold relative to the MALLET scores.
TABLE 4 HERE
We note that positioning and providing information are the most prevalent speech acts. The
mean raw MALLET-generated score for positioning is 0.349, indicating that the probability of
this speech act occurring is 34.9 percent. For providing information, that probability is 41.1
percent. Narrating is the third most frequently identified speech act, with a 15.6 percent
probability of being identified by the MALLET classifier. Requesting action and thanking are the
least common speech acts of members of Congress, representing 5.6 and 5.9 percent
probabilities of being identified by the classifier (and even less under the more conservative
estimates). All of these findings are consistent with those presented in Hemphill et al. (2013).
We highlight again that these speech acts are not mutually exclusive but may represent
combinations of speech acts. Presented in Table 5, the pairwise correlation analysis of both the
raw MALLET scores and the 0.50-threshold-based speech acts shows that narrating and
providing information and, separately, thanking and providing information are most strongly
correlated with each other. Given our interest in politicized framing on Twitter via conflated
speech acts, we also observe that the correlation coefficient for the positioning and providing
information speech acts is quite low, reflecting a very low association between them in terms of
both their raw MALLET scores and the 0.50 threshold measure.
TABLE 5 HERE
We examine the nature of hybrid speech acts and calculate them according to whether the
speech acts of interest met or did not meet the 0.50 threshold. That is, a speech act is
considered a “hybrid” if both of its constituent parts have raw MALLET scores of at least 0.50.
Presented in Table 6, 5.3 percent of the total sample of tweets (40,755 tweets) are a hybrid of
the narrating and providing information speech acts. Speech acts classified as a hybrid of
thanking and providing information represent 2.5 percent of the total sample (18,833 tweets).
Further, and contrasting with the relatively low correlation between them in Table 5, the
percentage of our sample represented by the positioning and providing information hybrid
speech act is 10.9 (83,396 tweets). The implication is that positioning and providing information
are not associated because the relationship between them is bifurcated: high levels of
positioning and low levels of providing information are negatively related, while low levels of
positioning and high levels of providing information are positively related. Whether these
patterns are connected to gender, chamber, and party-based differences is as of yet unknown
but can be determined with further analysis of all hybrid speech acts involving positioning and
providing information.
TABLE 6 HERE
Table 7 conveys the statistical output representing the first part of our analysis, focusing on how
non-hybrid speech acts are predicted by gender, chamber, and party. Given the cross-sectional
and longitudinal nature of the data and given the potential for individual-specific or time-specific
7error variance components to be significant, we conducted the Breusch-Pagan Lagrange
multiplier test and confirmed that a random effects model was appropriate. The results of these
analyses shows that gender is largely insignificant, that senators are more likely to position
while representatives are more likely to request action, and that Republicans request action
while Democrats provide information and thank. Narrating was not significantly different for any
of the independent variables. The lack of gender and party-based differences, particularly with
regard to positioning, contrasts strongly with existing research such as Russell (2017) and
Evans et al. (2014, 2016).
TABLE 7 HERE
The second part of our analysis focuses on hybrid speech acts, and we observe in the odds
ratios presented in Table 8 that gender-based differences are not at all significant, which again
contrasts with Evans et al. (2014, 2016). Chamber-based differences are also largely
insignificant except for the observation in model (7) that representatives are more likely to
engage in a providing information-thanking hybrid speech act. However, there are significant
differences between Democrats and Republicans across five of the seven hybrid forms
identified. There is also variance between parties in terms of preferences for specific hybrid
speech acts: Republicans are more likely to combine requests for action and positioning as well
as combine requests for action and providing information. Democrats are more likely to engage
in speech acts that combine positioning and providing information, positioning and thanking, and
providing information and thanking.
TABLE 8 HERE
These separate analyses of non-hybrid and hybrid speech acts have implications for how
framing occurs on Twitter. We observe in Table 3 that senators are more likely to engage in
positioning speech acts while representatives are more likely to engage in requests for action
and thanking speech acts. These patterns hold up in later findings on the hybrid forms (Table 8)
in that senators are more likely to simultaneously position and provide information while
representatives are more likely to provide information and thank. As a result, senators conflate
positioning and providing information much more than representatives.
Notable among our findings are the differences between Democrats and Republicans in terms
of how they frame their communications with hybrid speech acts. Statistically significant
differences based on random effects logistic regressions indicate that Democrats
position-provide information, position-thank, and provide information-thank more than
Republicans. Republicans, however, differ from Democrats in that they are more likely to
position-request action and provide information-request action. We also observe that
Republicans are no different from Democrats with regard to positioning or providing information
until they are conflated with requests for action. As such, requests for action are a distinctly
Republican framing strategy. At the same time, we observe that Democrats are significantly
more likely to engage in, separately, providing information and thanking. They are also more
likely to engage in these speech acts in their hybrid forms, implying that providing information
and thanking are crucial features for the Twitter-based Democratic framing strategy.
8Future Work
There are a number of ways that this line of research can be further expanded. One can expand
the nature of the hybrid speech act to include more than two speech acts at a time or,
alternatively, attend to the gradient of the raw MALLET score for positioning (and all other
speech acts) by comparing how these hybrid versions are predicted at more conservative
measurements. For example, a comparison could be made between 0.50 scores and 0.60 or
0.75 scores, the assumption being that more restrictive (higher) measures indicate stronger
positioning and thus potentially more polarizing rhetoric. One can also bridge speech acts with
other communication tools used by members of Congress on Twitter, the hashtag in particular.
Given connections identified between congressional framing of political issues via hashtags
(Hemphill, Culotta, & Heston, 2013), positioning and providing information speech acts can be
examined in the context of specific issues.
Future research can also attempt to build alternative automated classifiers using the scikit-learn
(Pedregosa et al., 2011) Python packages. Given the increased accessibility to powerful yet
easy-to-implement machine learning tools since 2013, we could, for example, compare our
maximum entropy classifier with a broad range of other techniques (e.g., logistic regression or
random forest models that provide transparent decisions on class assignments). The implication
is that there may be more computationally efficient and transparent classifiers. For instance,
logistic regression as a classifier is arguably a technique that is relatively better known and
understood by a broad audience; using logistic regression might thus make our results more
explainable and accessible. Given the increased concerns surrounding transparency and
accountability in data science, the use of more common techniques arguably offers benefits
over more computationally complex approaches in terms of explainability. Finally, we can
mitigate suspicions we have about the classifier with a more thorough comparison across
additional classifiers, whether they run in MALLET, Python, or another program or language.
Most importantly, the hybrid-focused analysis of speech acts among congressional Twitter users
has shown that, while both chambers and both parties are not significantly different from each
other in terms of how they engage in positioning speech acts, positioning-based hybrids are
distinctive framing tools for each party. Future research can help understand this in further detail
by examining, for example, whether a URL-based reference is included. More than 83K tweets
represent both positioning and providing information, but nearly 124K positioning tweets include
URLs according to our data. How these URL-included tweets are distinct from the hybrid tweet
is worth exploring.
Conclusion
We have determined that (a) information provision remains the most common speech act
among members of Congress on Twitter, (b) information provision and political positioning are
frequently conflated within hybrid tweets, and (c) females, senators, and Democrats are much
more likely to tweet. Together, these findings suggest that members of Congress have
developed routines of Twitter use that prioritize one-way communication and that attempt to
9subtly serve their framing efforts as they relate to positioning speech acts. How a particular
group of MCs uses Twitter at any given time seems to change -- e.g., whether women or
senators are more active -- but does not indicate dramatic advantages or disadvantages among
these groups in terms of access or impact.
A tension remains in U.S. political culture regarding the extent to which new technologies hinder
or facilitate political discussion online. Some are more optimistic (Delli Carpini, 2000), but we are
more cautious in light of the analysis above. We believe that positioning is an analogue to more
polarizing language than simply sharing information or making simple requests of constituents.
At stake is the potential for Twitter to limit democratic dialogue (Theocharis, Barberá, Fazekas,
Popa, & Parnet, 2016). This is possibly a premature concern given updates to how Twitter can
be used according to its guidelines4 as well as how politicians and the more politically active
public has modified and will continue to modify their online communications. Nevertheless, we
can offer suggestions for journalists and general citizens when encountering varying degrees of
information provision and positioning in congressional speech acts on Twitter.
First, recognizing the implications of Internet-based communications supplementing traditional
news outlets (Brainard, 2015), we point out that Twitter is not a passive activity. Increased use
by the general public with regarding to congressional tweets could in fact foster democratic
dialogue if constituents are providing clear and direct responses and commentary to their
elected officials. Journalists and citizens must actively engage members of Congress with
questions, feedback, and suggestions, and Twitter cannot be the only site of interaction. With
specific regard to Twitter, we know that this works as members of Congress are connected to
their constituents; i.e. the flow of information does in fact go both directions. However, we also
know that politicians are more influenced by those within their party as well as those that are
more politically engaged than by other users (Barberá et al., 2014). Which of these effects are
stronger -- co-partisanship or political engagement -- is in flux, and increases in Twitter-based
activity between members of Congress and their constituents will lead to a critical mass of
non-co-partisan-oriented dialogue. At that point, members of Congress would be forced to
address on Twitter the concerns of all constituents.
Second, savvy consumers of Congress’s tweets must be better informed regarding politicians’
attempts to blur the line between information provision and position-taking. For journalists, they
must revert back to more conservative methods and confirm the accuracy of every statement
posted by a politician on Twitter. This is particularly important for those tweets classified as
positioning and providing information speech acts, which we believe convey the most
newsworthy information. Staunching the flow of inaccurate information from politician to
journalist will have a significant impact on the news received by the general public, but it will not
affect citizens that receive and retweet information directly from their elected officials. For this
group, they must be made more literate regarding the potential for hybrid speech acts. One
simple but imperfect method is to verify any external links (i.e. URLs) provided in a politician’s
tweet. This will be a small but significant step toward establishing the credibility of the politicians’
4
https://help.twitter.com/en/rules-and-policies/twitter-rules
10statements and, in turn, potentially parsing out the constituents of a hybrid framing strategy on
Twitter.
References
Barberá, P., Bonneau, R., Egan, P., Jost, J. T., Nagler, J., & Tucker, J. (2014). Leaders or
followers? Measuring political responsiveness in the US Congress using social media data.
In 110th American Political Science Association Annual Meeting. Retrieved from
http://pablobarbera.com/static/barbera_twitter_responsiveness.pdf
Barberá, P., Bonneau, R., Jost, J. T., Nagler, J., & Tucker, J. (2013). Is There Anybody out
There? The Effects of Legislators’ Communication with their Constituents. In Text as Data
Workshop. London: London School of Economics and Political Science (pp. 1–24).
Citeseer.
Bell, A. (1984). Language style as audience design*. Language In Society, 13( 2), 145–204.
Brainard, C. (2015). The changing ecology of news and news organizations: implications for
environmental news. In A. Hansen & R. Cox (Eds.), The Routledge Handbook of
Environment and Communication. London and New York: Routledge.
Buckles, B. K. (2017). Keeping Constituents Informed? US House and Senate Armed Services
Committee Members’ Use of Twitter. Robert Morris University. Retrieved from
http://search.proquest.com/openview/fbfffee97acb9e1ec99ef6c6cf497f97/1?pq-origsite=gsc
holar&cbl=18750&diss=y
Chi, F., & Yang, N. (2011). Twitter adoption in Congress. Review of Network Economics, 10(1).
https://doi.org/10.2202/1446-9022.1255
Coddington, M., Molyneux, L., & Lawrence, R. G. (2014). Fact checking the campaign: how
political reporters use Twitter to set the record straight (or not). International Journal of
Press/Politics, 19(4), 391–409.
Cook, F. L., Tyler, T. R., Goetz, E. G., Gordon, M. T., Protess, D., Leff, D. R., & Molotch, H. L.
(1983). Media and agenda setting: effects of the public, interest group leaders, policy
makers, and policy. Public Opinion Quarterly, 47(1), 16–35.
Cunha, E., Magno, G., Comarela, G., Almeida, V., Gonçalves, M. A., & Benevenuto, F. (2011).
Analyzing the Dynamic Evolution of Hashtags on Twitter: A Language-based Approach. In
Proceedings of the Workshop on Languages in Social Media (pp. 58–65). Stroudsburg, PA,
USA: Association for Computational Linguistics.
Danescu-Niculescu-Mizil, C., Gamon, M., & Dumais, S. (2011). Mark My Words!: Linguistic
Style Accommodation in Social Media. In Proceedings of the 20th International Conference
on World Wide Web (pp. 745–754). New York, NY, USA: ACM.
Danescu-Niculescu-Mizil, C., West, R., Jurafsky, D., Leskovec, J., & Potts, C. (2013). No
Country for Old Members: User Lifecycle and Linguistic Change in Online Communities. In
Proceedings of the 22Nd International Conference on World Wide Web (pp. 307–318). New
York, NY, USA: ACM.
Delli Carpini, M. X. (2000). Gen.com: Youth, Civic Engagement, and the New Information
Environment. Political Communication, 17(4), 341–349.
Druckman, J. N. (2001). On the Limits of Framing Effects. The Journal of Politics, 63( 4),
1041–1066.
Druckman, J. N., Hennessy, C. L., Kifer, M. J., & Parkin, M. (2009). Issue Engagement on
11Congressional Candidate Web Sites, 2002—2006. Social Science Computer Review, 28( 1),
3–23.
Druckman, J. N., Kifer, M. J., & Parkin, M. (2017). Consistent and Cautious: Online
Congressional Campaigning in the Context of the 2016 Presidential Election. In J. C.
Baumgartner & T. L. Towner (Eds.), The Internet and the 2016 Presidential Campaign (pp.
3–24). Lexington Books.
Edwards, G. C., III, & Wood, B. D. (1999). Who Influences Whom? The President, Congress,
and the Media. The American Political Science Review, 93(2), 327–344.
Entman, R. M. (2007). Framing bias: media in the distribution of power. The Journal of
Communication, 57( 1), 163–173.
Evans, H. K., Cordova, V., & Sipole, S. (2014). Twitter Style: An Analysis of How House
Candidates Used Twitter in Their 2012 Campaigns. PS, Political Science & Politics, 47(2),
454–462.
Evans, H. K., Ovalle, J., & Green, S. (2016). Rockin’ robins: Do congresswomen rule the roost
in the Twittersphere? Journal of the Association for Information Science and Technology,
67(2), 268–275.
Frechette, C., & Ancu, M. (2017). Campaigning in 140 Characters: A Content Analysis of Twitter
Use by 2016 U.S. Congressional Candidates. In J. C. Baumgartner & T. L. Towner (Eds.),
The Internet and the 2016 Presidential Campaign (pp. 25–50). Lexington Books.
Gainous, J., & Wagner, K. M. (2014). Tweeting to Power: The Social Media Revolution in
American Politics. OUP USA.
Gentzkow, M., & Shapiro, J. M. (2011). Ideological segregation online and offline. The Quarterly
Journal of Economics, 1 26( 4), 1799–1839.
Giles, D. C. (2002). Parasocial Interaction: A Review of the Literature and a Model for Future
Research. Media Psychology, 4 (3), 279–305.
Giles, H., Coupland, J., & Coupland, N. (Eds.). (1991). Contexts of Accommodation:
Developments in Applied Sociolinguistics. Cambridge University Press.
Glassman, M. E., Straus, J. R., & Shogan, C. J. (2009). Social networking and constituent
communication: Member use of Twitter during a two-week period in the 111th Congress
(Report No. R40823). Washington, D.C. Retrieved from
http://assets.opencrs.com/rpts/R40823_20090921.pdf
Golbeck, J., Grimes, J. M., & Rogers, A. (2010). Twitter use by the U.S. Congress. Journal of
the American Society for Information Science and Technology , 61( 8), 1612–1621.
Graham, T., Jackson, D., & Broersma, M. (2014). New platform, old habits? Candidates’ use of
Twitter during the 2010 British and Dutch general election campaigns. New Media &
Society, 1461444814546728.
Granberg-Rademacker, J. S., & Parsneau, K. (2018). Tweet You Very Much: An Analysis of
Candidate Twitter Usage from the 2016 Iowa Caucus to Super Tuesday. In C. J. Galdieri, J.
C. Lucas, & T. S. Sisco (Eds.), The Role of Twitter in the 2016 US Election (pp. 21–44).
Cham: Springer International Publishing.
Haber, S. (2011). The 2010 US Senate Elections in 140 Characters or Less: An Analysis of How
Candidates Use Twitter as a Campaign Tool. Retrieved from
https://auislandora.wrlc.org/islandora/object/1011capstones%3A154/datastream/PDF/view
Hamby, P. (2013). Did Twitter kill the boys on the bus? Searching for a better way to cover a
campaign. Discussion Paper Series #D-80 Joan Shorestein Center on the Press, Politics
and Public Policy.
Hemphill, L., Culotta, A., & Heston, M. (2013). Framing in social media: how the U.S. Congress
uses Twitter hashtags to frame political issues.
12Hemphill, L., Culotta, A., & Heston, M. (2016). #Polar Scores: Measuring Partisanship Using
Social Media Content. Journal of Information Technology & Politics, 13( 4), 365–377.
Hemphill, L., Otterbacher, J., & Shapiro, M. (2013). What’s congress doing on twitter? In
Proceedings of the 2013 conference on Computer supported cooperative work (pp.
877–886). New York, NY, USA: ACM.
Huang, J., Thornton, K. M., & Efthimiadis, E. N. (2010). Conversational Tagging in Twitter. In
Proceedings of the 21st ACM Conference on Hypertext and Hypermedia (pp. 173–178).
New York, NY, USA: ACM.
Hu, Y., Talamadupula, K., & Kambhampati, S. (2013). Dude, srsly?: The Surprisingly Formal
Nature of Twitter’s Language. In Seventh International AAAI Conference on Weblogs and
Social Media. Retrieved from
https://www.aaai.org/ocs/index.php/ICWSM/ICWSM13/paper/viewPaper/6139
Karlsen, R. (2015). Followers are opinion leaders: The role of people in the flow of political
communication on and beyond social networking sites. European Journal of Disorders of
Communication: The Journal of the College of Speech and Language Therapists, London,
30(3), 301–318.
Kedrowski, K. M. (2000). How members of Congress use the media to influence public policy. In
D. A. Graber (Ed.), Media Power in Politics. Washington, D.C.: Congressional Quarterly,
Inc.
Lee, H. S. (2009). News Media Conditions Presidential Responsiveness to the Public. In APSA
Annual Conference. Toronto.
Lieber, P. S., & Golan, G. J. (2011). Political public relations, news management, and agenda
indexing. In J. Stromback & S. Kiousis (Eds.), Political Public Relations: Principles and
applications. New York: Routledge.
Marwick, A. E., & Boyd, D. (2010). I tweet honestly, I tweet passionately: Twitter users, context
collapse, and the imagined audience. New Media & Society, 13(1), 114–133.
McCallum, A. K. (2002). MALLET: A Machine Learning for Language Toolkit. Retrieved from
http://mallet.cs.umass.edu
McMillan, S. J. (2002). Exploring Models of Interactivity from Multiple Research Traditions:
Users, Documents, and Systems. In L. A. Lievrouw & S. Livingstone (Eds.), Handbook of
New Media: Social Shaping and Consequences of ICTs (pp. 163–182). 1 Oliver’s Yard, 55
City Road London EC1Y 1SP : SAGE Publications, Ltd.
Nigam, K., Lafferty, J., & Mccallum, A. (1999). Using Maximum Entropy for Text Classification.
Science, 1, 61–67.
Niven, D., & Zilber, J. (2001). Do Women and Men in Congress Cultivate Different Images?
Evidence from Congressional Web Sites. Political Communication, 18( 4), 395–405.
Norris, P., & Curtice, J. (2008). Getting the Message Out: A Two-Step Model of the Role of the
Internet in Campaign Communication Flows During the 2005 British General Election.
Journal of Information Technology & Politics, 4( 4), 3–13.
Parmelee, J. H. (2013). The agenda-building function of political tweets. New Media & Society,
16(3), 434–450.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., … Duchesnay, É.
(2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research:
JMLR, 12( Oct), 2825–2830.
Russell, A. (2017). U.S. Senators on Twitter: Asymmetric Party Rhetoric in 140 Characters.
American Politics Research, 1532673X17715619.
Shapiro, M. A., & Hemphill, L. (2017). Politicians and the Policy Agenda: Does Use of Twitter by
the U.S. Congress Direct New York Times Content? Policy & Internet, 9( 1), 109–132.
13Straus, J. R., Glassman, M. E., Shogan, C. J., & Smelcer, S. N. (2013). Communicating in 140
Characters or Less: Congressional Adoption of Twitter in the 111th Congress. PS, Political
Science & Politics, 46(1), 60–66.
Stromer-Galley, J. (2000). On-Line Interaction and Why Candidates Avoid It. The Journal of
Communication, 50( 4), 111–132.
Theocharis, Y., Barberá, P., Fazekas, Z., Popa, S. A., & Parnet, O. (2016). A Bad Workman
Blames His Tweets: The Consequences of Citizens’ Uncivil Twitter Use When Interacting
With Party Candidates. The Journal of Communication, 66( 6), 1007–1031.
Verweij, P. (2011). Twitter links between politicians and journalists. Journalism Practice, 6(5-6),
680–691.
Williams, C. B., & Gulati, G. J. (2010). Communicating with constituents in 140 characters or
less: Twitter and the diffusion of technology innovation in the United States Congress.
SSRN eLibrary. Retrieved from http://ssrn.com/paper=1817053
14Table 1. Definitions and examples of Twitter-based speech acts
Code Definition Example
Narrating Telling a story about their day, “headed up to the Fox News camera for an
describing activities interview” (Rep. Ron Paul, R-TX)
Positioning Situating one's self in relation to “A9: Theoretically, not realistically. HC
another politician or political issue, spending is growing 4x inflation and
may be implied rather than explicit driving our debt. Let’s tackle the real
threat. #ryanttv” (Rep. Paul Ryan, R-WI)
Providing Pointing to a resource URL, telling “Harkin Announces More Than $300,000
information you where you can get more info for Housing in Tama County
http://1.usa.gov/lf6Aem” (Sen. Tom Harkin,
D-IA)
Requesting Explicitly telling followers to go do “RSVP to my Immigration Forum with Rep.
action something online or in person (not Luis Gutierrez this Saturday in Brooklyn
just visiting a link but asking them http://t.co/qTcWugs” (Rep. Yvette Clark,
to do something like sign a petition, D-NY)
apply, vote) - look for action verbs
Thanking Says nice things about or thanks “@rmartindc Thanks. MoC's handwriting is
someone else, e.g. congratulations, probably on par with M.D.'s. Glad I could
compliments make your job easier.” (Rep. John
Shimkus, R-IL)
Note: Based on Hemphill et al. (2013), Table 1.
15Table 2. Mean classification accuracy on test data: 10-fold cross-validation procedure
Classifier Narrating Positioning Providing info Req. action Thanking
Maximum 0.83 0.71 0.91 0.91 0.95
Entropy (-0.05) (-0.06) (-0.03) (-0.03) (-0.03)
16Table 3. Predicting frequency of Twitter posts, OLS
Frequency of tweet
Male -170.713**
(66.660)
Senate 354.935***
(66.681)
Republican -159.685***
(54.839)
Constant 837.300***
(59.602)
F 16.32***
R2 0.107
N 414
17Table 4. Descriptive statistics for speech acts, two measures
Raw MALLET scores Applying 0.50 threshold
Mean Standard Dev. Mean Standard Dev.
Narrating 0.156 0.201 0.083 0.276
Positioning 0.349 0.267 0.274 0.446
Providing info 0.411 0.395 0.410 0.492
Requesting action 0.056 0.101 0.014 0.116
Thanking, etc. 0.059 0.126 0.026 0.158
Note: “Raw MALLET scores” refer to the probability that an individual tweet belongs in a given
class; “Applying 0.50 threshold” refers to the proportion of tweets that are assigned probabilities
above 0.50 for a given class.
18Table 5. Pairwise correlation coefficients of speech acts, raw MALLET scores / 0.50 threshold
Narrating Positioning Providing info Req. action Thanking
Narrating 1/1
Positioning -0.162 /-0.063 1/1
Providing info 0.259 / 0.141 -0.048 /-0.017 1/1
Req. action 0.083 / 0.067 -0.216 /-0.061 0.018 / 0.090 1/1
Thanking -0.021 /-0.030 -0.165 /-0.080 0.398 / 0.180 -0.076 /-0.017 1/1
19Table 6. Descriptive statistics for hybrid speech acts
Mean Standard Dev.
Positioning & Narrating 0.015 0.122
Positioning & Providing info 0.109 0.312
Positioning & Requesting action 0.001 0.024
Positioning & Thanking 0.001 0.038
Providing info & Narrating 0.053 0.224
Providing info & Requesting action 0.006 0.078
Providing info & Thanking 0.025 0.155
Note: “Mean” refers to the proportion of tweets that are assigned probabilities above 0.50 for
each of the two classes listed.
20Table 7. Predicting speech acts, generalized least squares
(1) (2) (3) (4) (5)
Narrating Positioning Providing info Req. action Thanking
Male 0.002 -0.004 0.002 0.003* 0.000
(0.004) (0.006) (0.009) (0.001) (0.002)
Senate -0.001 0.020*** -0.006 -0.004** -0.006**
(0.005) (0.006) (0.012) (0.002) (0.002)
Republican 0.001 -0.006 -0.041*** 0.010*** -0.009***
(0.003) (0.005) (0.009) (0.001) (0.002)
Constant 0.158*** 0.347*** 0.439*** 0.049*** 0.065***
(0.003) (0.005) (0.010) (0.001) (0.002)
Model type Random Random Random Random Random
effects effects effects effects effects
Χ2 0.56 14.19*** 21.66*** 82.13*** 29.72***
Groups 415 415 415 415 415
N 765,826 765,826 765,826 765,826 765,826
0.10, pTable 8. Predicting hybrid speech acts, logistic
(1) (2) (3) (4) (5) (6) (7)
Positioning Positioning Positioning Positioning Providing Providing Providing
& Narrating & Providing & Req. & Thanking info & info & info &
info action Narrating Req. Thanking
action
Male 0.994 0.943 0.992 1.035 1.045 1.175 1.021
(0.065) (0.052) (0.142) (0.105) (0.061) (0.120) (0.078)
Senate 1.070 1.093* 0.843 1.052 0.937 0.906 0.740***
(0.069) (0.060) (0.116) (0.104) (0.054) (0.090) (0.056)
Repub. 0.984 0.789*** 1.840*** 0.673*** 0.980 1.615*** 0.852**
(0.053) (0.036) (0.223) (0.058) (0.047) (0.136) (0.054)
Constant 0.014*** 0.132*** 0.000*** 0.001*** 0.053*** 0.003*** 0.025***
(0.001) (0.007) (0.000) (0.000) (0.003) (0.000) (0.002)
Model Random Random Random Random Random Random Random
type effects effects effects effects effects effects effects
Χ2 1.22 36.78*** 29.24*** 22.33*** 1.82 45.24*** 22.42***
Groups 415 415 415 415 415 415 415
N 765,826 765,826 765,826 765,826 765,826 765,826 765,826
Note: Odds ratios are presented; standard errors in parentheses; *, **, *** represent pYou can also read
